Employee types and endogenous organizational
design: An experiment
Antoni Cunyaty
University of Valencia
Randolph Sloofz
University of Amsterdam and Tinbergen Institute
November 19, 2007
Abstract
When managers are su¢ ciently guided by social preferences, incentive provision through an organizational mode based on informal implicit
contracts may provide a cost-e¤ective alternative to a more formal mode
based on explicit contracts and monitoring. This paper reports the results
from a laboratory experiment designed to test whether organizations make
e¤ective use of the available preference types within their work force when
drafting their organizational design. Our main …nding is that they do not
do so; the signi…cant impact managers’preferences have on the behavior
of workers in the organization seems to be overlooked by those choosing
the organizational mode.
1
Introduction
A major research theme within organizational economics is how to motivate
employees to exert well-directed e¤ort. This issue is typically addressed using
the principal-agent model as point of departure. In the standard version of this
model the agent is assumed to solely care about his own monetary compensation
and to dislike e¤ort. Similarly so, the principal just wants to maximize her
own net pro…t and does not care about the agent’s well-being. Given these
We would like to thank Enrique Fatás and LINEEX for their hospitality while we conducted the experiment. Programming support of Héctor Solaz is gratefully acknowledged.
Antoni Cunyat acknowledges …nancial support from the Spanish Ministry of Education and
Science, Project SEJ 2005-08054/ECON and from Conselleria d’Empresa, Universitat i Ciència of Valencian Government through project GV2006-201.
y Departament d’Anàlisi Econòmica.
Av/ dels Tarongers s/n.
Campus dels
Tarongers. Edi…ci Departamental Oriental. 46022 Valencia (Spain). Antonio.Cunat@uv.es.
http://www.uv.es/acunat.
z School of Economics. Roetersstraat 11.
1018 WB Amsterdam.
the Netherlands.
r.sloof@uva.nl.
1
assumptions it is derived how monetary incentives should be optimally designed
in order to motivate the agent to put in su¢ cient e¤ort.
Many empirical studies have found, however, that people may have alternative motivations that go beyond material self-interest. Fairness, altruism,
empathy and a preference to react in kind to kind or unkind actions of others (reciprocity) are among the various alternative motivations identi…ed. The
presence of such ‘social preferences’ may have profound implications for the
provision of e¤ort incentives. In the context of the principal-agent relationship,
for instance, workers are more easily persuaded to exert e¤ort when they know
that their manager cares about their well-being and thus will reward higher
e¤ort with a larger (non-contractible) bonus. E¤ort levels will then be higher in
equilibrium, thereby increasing e¢ ciency. For this reason it may actually pay
for …rms to select and hire ‘empathic’managers who do not solely care about
pro…t maximization; their personality type (i.e. preferences) helps in overcoming a di¢ cult incentive problem with the workers (cf. Rotemberg and Saloner
(1993), Rotemberg (1994), and Hermalin (2001, Section 4.2)).1 In the same
spirit, …rms may want assign particular preference types within their work force
to jobs where these social preferences are most e¤ective.
In this paper we leave aside the question of screening and selection of employees on the basis of social preferences. Rather, we intend to investigate whether
organizations will make use of the existing social preferences within their work
force, by choosing an organizational design that is particular conducive to their
e¤ectiveness. Because it is notoriously di¢ cult to gather …eld data on this, we
make use of laboratory experiments to test the relevant theoretical predictions
at hand.
In our experiment we simplify matters by assuming that there are two types
of organizational modes, each corresponding to a di¤erent role (’leadership
style’) for managers. In the …rst type managers are hired to inspire and to
motivate the work force. Rather than implementing a formal system that relies
on explicit incentive contracts and active monitoring, managers should instill
and maintain a culture that hard work will be rewarded by the organization.
This implicit contract then substitutes for a more costly explicit performance
measurement and evaluation system. We represent this particular organizational mode in highly reduced form with the motivation game M depicted in
Figure 1a below. This game corresponds to the trust game used by Kreps (1990)
to model corporate culture. It also represents a simpli…ed version of the game
used by Rotemberg and Saloner (1993) to study the impact of leadership style
on workers’incentives to innovate.2
1 Rotemberg and Saloner (2000) provide a model where CEOs have a vision, which means
that they are consistently biased toward certain kind of projects and against others. This may
ameliorate incentive problems within the organization, i.e. workers are led to work hard on
projects toward which the CEO is biased. Like a preference for empathy, CEO biasedness in
a sense thus may work as a commitment device.
2 In spirit the M -game corresponds to the ’loose supervision’regime in the model of supervision and workgroup identity studied by Akerlof and Kranton (2005). The inspection game
I (to be discussed shortly) then corresponds to their ’strict supervision’regime.
2
< Insert Figure 1 about here >
In game M a worker …rst decides whether to put in (high) e¤ort or to shirk.
In case the worker shirks, it is assumed that he does not get a reward (on top of
his wage). If, however, the worker exerts e¤ort, the manager decides whether to
reward him with a bonus or not. This bonus is not part of an enforceable payfor-performance contract though. A sel…sh manager will therefore not pay the
bonus and, anticipating this, the worker will not put in high e¤ort. If, however,
the manager could credibly commit to pay the bonus (only) when the worker
exerts e¤ort, the worker would be motivated to do so.
The second type of organizational mode relies more strongly on formal contracts and explicit monitoring. Here managers are hired to supervise and to
monitor workers. Workers receive a given wage for putting in e¤ort, but are
…ned or …red whenever they are caught shirking. In line with Calvo and Wellisz
(1978), the role of the manager is then to supervise, i.e. to check whether the
worker does not deliver substandard work. We model this particular situation
with the inspection game I depicted in Figure 1b, for high (IH ) and low (IL )
inspection costs separately. Here the manager (as a supervisor) …rst decides
whether to monitor or not. If she does so, the only (relevant) option for the
worker is to put in high e¤ort. But if the manager decides not to monitor, the
worker decides whether to shirk or to exert (high) e¤ort. Payo¤s are such that
sel…sh workers will shirk if not monitored. Realizing this, the manager commits
to monitor in the …rst stage.3
Assuming sel…sh preferences, the worker is predicted to exert e¤ort under
organizational structure I whereas under structure M he does not. Structure I
is thus more e¢ cient, and therefore likely to be preferred by the owner of the
…rm (assuming that she shares in the e¢ ciency gains). Things change when
employees may have social preferences. Putting a very empathic (or reciprocal)
employee in the manager’s position under structure M will then yield …rst best.
This holds because, anticipating that he will be rewarded with a bonus by the
empathic manager, the worker puts in e¤ort and does not shirk. And compared
to structure I, organizational mode M saves on the costs of the monitoring
technology.4 Structure M thus becomes relatively more attractive the more
empathic the employees are. Another intuitive prediction is that the less costly
the inspection system under structure I is, the more it takes for the owner to
3 Game I re‡ects a simpli…ed version of an inspection game where the manager can commit
to a particular inspection strategy. In a more general setup, the manager commits to a
particular inspection probability, such that the worker is just induced to exert e¤ort with
probability one; see Section 5 in Avenhaus, von Stengel, and Zamir (2002) for a full discussion
and justi…cation of this game. In their real e¤ort experiment Dickinson and Villeval (2004)
also use a inspection game in which the principal/manager can commit ex ante to a given
monitoring technology. For simplicity, here we restrict the manager to just two inspection
probabilities, either zero (no monitoring at all) or one (always monitor).
4 This explains why the maximum joint payo¤s for the worker and the manager under M
(1100) are higher than the maximum joint payo¤s under I (980). In the Appendix we provide
an elaborate justi…cation of the particular parameterization depicted in Figure 1 (and used in
the experiment).
3
prefer the M structure. Lower inspection costs can e¤ectively be represented in
Figure 1b by having payo¤s of 440 (instead of 360) after the manager’s decision
to monitor. More generally, theory predicts that the M structure which relies
on empathic managers becomes less likely the more cost-e¤ective the formal
performance measurement system under mode I is.
The experiment that we use to test the above predictions consists of two
parts. In part one subjects make decisions in 9 di¤erent games that all have
the same structure as those in Figure 1. That is, a …rst mover (player A) …rst
decides whether to ‘stay-out’or to ‘enter’. Only if player A enters, player B is
called on to choose between left and right. Subjects both take decisions in the
player A role and (conditional) decisions in the player B role. In that way the
…rst part generates an individual ‘track record’(ei ,ri ), with ei (ri ) the number
of entry decisions (right choices) the subject made. This individual track record
can be seen as a (imprecise) measure of the subject’s preference type.
In part two subjects …rst learn their role, either owner of a …rm or employee.
Roles are kept …xed during all 15 periods of this part. In the …rst …ve periods the
organizational mode is exogenously given, by game M say. In each period …rms
consisting of one owner and two employees are formed, based on a strangers
design. All …rm members observe the track records of the two employees (but
not of the owner). Based on this information, the owner decides who to put in
the manager’s (i.e. second mover) position. After that the two employees play
the corresponding game, yielding payo¤s to them as re‡ected in Figure 1a. The
owner’s payo¤s equal those of the manager (and are private information).
The next …ve periods consider the other organizational mode, here game I
of Figure 1b. The sequence of events is the same as in periods 1 to 5. Moreover,
also in game I the owner’s payo¤s correspond with those of the manager, which
is now the …rst mover in this game. In the …nal …ve periods the organizational
mode is made endogenous and owners …rst choose between game M and game
I, before they assign their workers to a particular position. Overall four sessions
are conducted, that di¤er in the order in which the two organizational modes
are played and in the version of the inspection game considered (IL or IH ).
Our main …ndings can be summarized as follows. The informational value
of subjects’ track records lies in their ri -values. In particular, in both games
employees with the …rst mover role are more likely to ‘enter’ the higher the
ri value is of the second mover with whom they are matched. So indeed, the
more ‘empathic’ second movers are, the higher the willingness of …rst movers
to enter the reciprocal relationship. Firm owners seem to overlook this mechanism, however, when assigning their employees to di¤erent roles. They naively
assume that employees’decisions are mostly a¤ected by their own track record
characteristics only. As a result, in the M -game they typically assign the role
of manager (second mover) to the employee with the lower ri -value in his track
record, i.e. to the more sel…sh employee. These assignments appear suboptimal,
because pro…t would have been higher if they would assign roles the other way
around. Also the choice between di¤erent organizational modes is not guided
by employees’ ri values (as they should). Overall we therefore conclude that
owners in our experiment do not make e¤ective use of the social preferences
4
within their work force.
Numerous laboratory experiments have already been conducted that relate
to the above discussed issues of endogenous organizational design. In a series of
papers, for instance, Ernst Fehr and various coauthors have studied the choice
of optimal incentive contracts. A main common …nding is that social preferences can serve as a useful (i.e. cost-e¤ective) contract enforcement device and
contracts may therefore be deliberately left incomplete.5 In the terminology
used before, in practice …rms may prefer implicit contracts over explicit incentive contracts, although under sel…sh preferences the latter would be optimal.
Similarly so, others have explicitly studied the e¤ect of monitoring on behavior in more detail; see e.g. Dickinson and Villeval (2004) and Schweitzer and
Ho (2005). Compared to these previous experiments, we study (implicit and
explicit) contracts and monitoring in highly reduced form (cf. Figure 1). The
main contribution of our experiment is that we explicitly relate these (reduced
form) organizational choices to the observed characteristics (‘track records’) of
the employees that are to be a¤ected by these instruments. Apart from that,
we consider the endogenous allocation of roles within organizations.
This paper proceeds as follows. Assuming that employees may care about
the well-being of others, we derive in the next section the formal predictions
and hypotheses that are put to the test. Section 3 presents the details of our
experimental design whereas Section 4 reports the results. The …nal section
summarizes and concludes.
2
Theoretical predictions and hypotheses
Our experiment is based on the M -game and the I-game as depicted in Figure
1.6 Both games have the same general decision structure, which is re‡ected in
Figure 2 below. Player A …rst chooses between Stay Out and Enter. If player
A enters, player B subsequently chooses between Left and Right. Payo¤s are
such that choosing left yields B the most in monetary payo¤s, whereas right
corresponds to sacri…cing to reward A for the ’kind’ choice to enter. From
d > c > b > a it immediately follows that, If both players are sel…sh, (Out,
Left) is the unique subgame perfect equilibrium.
< Insert Figure 2 about here >
Social preferences may lead players away from the ine¢ cient (Out,Left) outcome. Various alternative motivations have been identi…ed in the literature –
like fairness, altruism, empathy and reciprocity – and a number of theoretical models have been developed to capture these types of social preferences in
5 See e.g. Fehr et al. (1997, 2007), Fehr and List (2004) and Fehr and Schmidt (2000,
2004).
6 In the Appendix we discuss a basic reduced form model of endogenous organizational
design that underlies these two speci…c games.
5
a formal way. Prominent examples include Fehr and Schmidt (1999)´s model
of inequality-aversion, Charness and Rabin (2002)´s model of quasi-maximin
(social welfare) preferences, and Rabin (1993)´s model of intention-based reciprocity (see also Dufwenberg and Kirchsteiger (2004)). Although these models
can lead to quite di¤erent predictions in particular situations, a common theme
they share is that social preferences may be e¢ ciency enhancing. It is this
common aspect that we want to emphasize here.7
To illustrate the impact alternative motivations may have, we capture social
preferences in a very simple and stylized way. Let i and j denote player
i’s and j’s monetary payo¤s. Following Charness and Rabin (2002), we assume
that player i’s preferences take the following form (with i 6= j and i; j 2 fA; Bg):
Ui ( i ;
j)
=
=
i
i
+ (1
j + (1
j
i)
i
i)
i
if
if
i
i
>
j
(1)
j
In this speci…cation, parameter i gives the weight player i attaches to the
other player’s payo¤s when she herself is ahead. Parameter i re‡ects the corresponding weight when she is behind. Without any restrictions on i and i
utility function (1) can capture a range of di¤erent motivations. Charness and
Rabin (2002) use the results of a variety of simple games with a similar decision structure as in Figure 2, to estimate the values of i and i :They …nd
that on average players do not care about other players’payo¤s when they are
behind, but put a positive weight on the well-being of others when they are
ahead. In line with their estimates we therefore assume that 0 < i < 1 and
that i
0. These assumptions incorporate the inequality-aversion model of
Fehr and Schmidt (1999), which corresponds to i < 0 < i < 1 and j i j
i.
They are also in line with Hermalin (2001, Section 4.2), who assumes that a
player su¤ers from ’remorse’only if he is ahead, because of the unfairness of the
allocation.
Assuming preferences as in (1), the game is again easily solved by backward
d c 8
induction. Player B will choose Right whenever B
d a . Anticipating
this, player A enters only when B exceeds this threshold. Hence the predicted
outcome is Out when B <
and (In, Right) in case B
. (Outcome
(In,Left) is never observed on the equilibrium path.) This establishes that when
player B cares su¢ ciently about A’s well-being, the ine¢ cient outcome Out is
avoided. Note that i is the key parameter here. Following Rotemberg and
Saloner (1993) we say that a player is more empathic the higher his i is.
To capture the role of the owner of the …rm, next assume there is a third
player C who decides on role assignment. In particular, player C has two em7 Our experiment thus neither should be taken as an attempt to discriminate between
various types of social preferences, nor as providing a test of a particular version of social
preferences per se. Although in this section we use quasi-maximin preferences to derive and
illustrate the main implications in a parsimonious way, similar predictions would have been
obtained under relevant alternative speci…cations. For instance, incorporating Dufwenberg
and Kirchsteiger (2004) style intention-based reciprocity motivations leads to qualitatively
the same predictions.
8 Here we assume that in the knife-edge case where player B is indi¤erent (
), he
B =
chooses Right for sure.
6
ployees at her disposal, numbered such that 1
2 . Her task is to decide which
employee gets role A and who gets role B. In contrast to her two employees,
player C is assumed to be sel…sh. Her monetary payo¤s equal some weighted
combination A A +(1 A ) B (with A 2 [0; 1]) of the monetary payo¤s of A
and B;9 e¤ectively, player C cares about e¢ ciency. Then, knowing that in equilibrium the outcome either equals Out or (In, Right), player C prefers to assign
role B to the more empathic employee 2 . This follows because this maximizes
the probability that the more e¢ cient outcome (In,Right) is obtained.10
Apart from role assignment, player C may possibly also choose between game
M and game I. These two games share the same decision structure of Figure 2,
but payo¤s are di¤erent. In particular, bI > bM and cI < cM (cf. Figure 1).11
Moreover, in game M player C gets the same as player B (so A = 0) whereas in
game I her monetary payo¤s equal those of player A ( A = 1). The idea here is
that the owner C gets the same as the manager, which corresponds to player B
in the M -game and player A in the I-game. Given this payo¤ structure, player
C will choose game M only when outcome (In,Right) is expected in that game.
This yields her cM whereas game I gives her at most cI . If, however, outcome
Out would result in game M , player C prefers to choose game I. The latter
yields her at least bI ; which exceeds bM .
The predicted outcome within each game depends on how the player B’s
level of empathy B compares to the critical value
for that game, viz. either
12
player
C
is
best
o¤
by
choosing
game M (and putting
.
If
>
or
2
M
I
M
employee 2 in position B). Otherwise, C is better o¤ in the I-game. Hence we
obtain that C is more likely to choose game M the more empathic her employees
are.
Finally, we analyze the role of inspection costs in the I-game. Higher inspection costs correspond to lower values for bI . If player C would be perfectly
informed on her employees’empathy parameters i , variations in bI would not
a¤ect the predicted outcomes (as long as bI > bM ). However, it seems reasonable to assume that parameter i is employee i’s private information. The
owner may have a good idea about what the value of i is, but is not completely
sure about its exact value. In particular, the owner believes that i is drawn
from a particular probability distribution. If she chooses the I-game then, her
expected payo¤s equal a weighted average of bI and cI : If she opts for the M game instead she gets a weighted average of bM and cM : Other things equal,
the expected payo¤s of choosing the I-game are lower the lower bI is. Player C
is thus more likely to choose M over IH than over IL (cf. Figure 1).
9 In the experiment player C’s payo¤s are private information to her, so A and B cannot
be guided by empathic feelings towards player C. This justi…es that (1) does not include C .
1 0 Note that e¤ectively role assignment makes a di¤erence only when
< 2 . Then
1 <
the outcome is (In, Right) when employee 2 is put in the B-role and Out when employee 1 is
put in the B-role. Because c > b player C prefers the former assignment.
1 1 As explained in the Appendix, the …rst inequality b > b
I
M re‡ects the idea that the value
of the worker’s e¤ort exceeds the overall costs of a formal monitoring system. Restriction
cI < cM derives from the fact that installing a monitoring technology brings about (…xed)
investment costs, even when it is not actively used in the end.
1
1
1 2 For the parameters chosen in our experiment (cf. Figure 1) we have
M = 3 and I = 6 .
7
Taken together, we obtain the following …ve predictions that are put to the
test:
1. The more empathic player B is (i.e. the higher
is that he chooses Right as a second mover;
B
is), the more likely it
2. The more empathic player B is (i.e. the higher
is that player A enters;
B
is), the more likely it
3. Players C will assign the role of player B to the more empathic employee
within her work force (i.e. to the one with the higher i -value);
4. The more empathic the work force is (i.e. the higher 2 ), the more likely
it is that player C prefers the M -game over the I-game;
5. Ceteris paribus, player C is more likely to choose M over IH than over IL .
These formal predictions correspond to the intuition predictions discussed
in the Introduction. To see this, recall that in game M the worker moves …rst
whereas the manager moves second. The …rst prediction then reads: the more
empathic the manager is, the more likely it becomes that he will reward high
e¤ort with a bonus. As a result, the worker has a stronger incentive to put
in high e¤ort (cf. the second prediction). In contrast, in game I the manager
moves …rst and the worker moves second. Then the more empathic the worker is,
the more likely it is that he chooses high e¤ort. Anticipating this, the manager
is more likely to abstain from costly monitoring.
Turning to owner behavior, the third prediction implies that in the inspection mode, the more sel…sh employees within the …rm are the managers. In
contrast, in the motivation mode the most empathic types are assigned the role
of manager. Managers in the inspection mode will therefore on average be more
sel…sh than managers in the motivation mode. When owners can choose between organizational modes, we expect that they will opt for the motivation
game M only when they have at least one very empathic type among their employees. This employee will then be given the role of manager. Otherwise the
inspection game will be chosen (and the least empathic type will be given the
managerial position). Finally, the more costly the formal monitoring system is
(i.e. structure IH versus IL ), the more likely it is that structure M is chosen.
Like in reality, in the experiment employees do not observe the level of
empathy of their colleagues precisely, and neither do so owners C. Based on an
observable track record of past choices, however, an estimate ri of a player’s
empathy level i can be obtained. The formal hypotheses we want to test then
correspond to predictions 1 through 5 above, where i is replaced by its estimate
ri . Exactly how individual track records are generated in the experiment is
explained in the next section.
8
Table 1: Overview of sessions and treatments (in part 2)
session rounds 1 5 rounds 6 10 rounds 11 15
1
IL
M
IL versus M
2
M
IL
M versus IL
3
IH
M
IH versus M
4
M
IH
M versus IH
3
Experimental design
Our experiment is based on a 2 by 3 treatments design. In each session we
kept the two types of games (game M and game I) …xed. Between sessions
we varied the particular version of the inspection game, having either the one
representing low inspection costs (IL ) or the other one with high inspection costs
(IH ). We ran four sessions in total, which di¤ered according to (the order of) the
treatments considered. Table 1 provides an overview. All sessions were run in
May 2007 at the LINEEX laboratory of the University of Valencia. Overall 180
subjects participated, with 45 subjects per session. The subject pool consisted
of undergraduate students at the University of Valencia. The vast majority of
them (88%) were students in Economics or Business, 57% were male. They
earned on average 24:5 euros in somewhat less than 2 hours, including a show
up fee of 7 euros.
Each sessions consisted of two parts. At the beginning of the experiment
subjects were informed about this. They were also informed that possibly some
of the choices they made in part 1 would become observable to some other
participants in part 2.13 In particular, the instructions for part one explained
that:
“...It may happen that in part two some other participants get
some information about your decisions in part one. It is also possible
though that none of your part one decisions will ever become known
to any other participant. ...”14
Apart from this information, subjects were kept ignorant about the actual
content of part 2 until that part actually started.15
In the …rst part subjects made decisions for a series of nine extensive form
games that all have the same setup as in Figure 2. We used a neutral frame
1 3 In fact, the probability that none of a subject’s part 1 choices would ever become observable to any other participant in part 2 equalled 13 for each subject (see below).
1 4 This announcement has the clear disadvantage that it may in‡uence subjects’ decisions
in part 1. We considered it necessary though, in order to avoid any potential impression of
deception. Moreover, if we would not make the announcement, subjects would be surprised
at the start of part 2 when their choices of part 1 became known, and might think that it is
quite likely that another “surprise” will follow. This might then a¤ect their behavior in part
2.
1 5 The experiment was conducted in Spanish. An English translation of the instructions can
be found at the following url: ??
9
game
I
II
III
IV
V
VI
VII
VIII
IX
Table 2: Overview of the games in part one
A stays out
If A enters, B chooses
sacri…ce
(b; b)
(a; d) vs. (c; c)
d c
(1400,1400) (950,3650) vs. (2750,2750)
900
(1800,1800) (200,2900) vs. (2450,2450)
450
(2200,2200) (200,2900) vs. (2450,2450)
450
(1400,1400) (200,2900) vs. (2450,2450)
450
(1800,1800) (950,3650) vs. (2750,2750)
900
(2200,2200) (950,3650) vs. (2750,2750)
900
(1400,1400) (1250,3950) vs. (2600,2600)
1350
(1800,1800) (1250,3950) vs. (2600,2600)
1350
(2200,2200) (1250,3950) vs. (2600,2600)
1350
Remark: It holds that
=
sacrif ice
sacrif ice+reward
=
reward
c a
1800
2250
2250
2250
1800
1800
1350
1350
1350
0.33
0.16
0.16
0.16
0.33
0.33
0.5
0.5
0.5
(d c)
.
(d c)+(c a)
for the entire experiment, with A’s choosing between A1 (Stay Out) and A2
(Enter) and B’s choosing between B1 (Left) and B2 (Right). Table 2 provides
an overview of the games used. All subjects made choices for both roles, in all
nine games.16 Overall they thus made 18 choices in part 1. They were informed
that at the end of the experiment one of these choices was randomly selected
and paid (see below for more on this).
The nine di¤erent games of part 1 have been chosen as follows. The …rst
three games are just upscaled versions (by a factor of …ve) of the M -game, the
IH -game and the IL -game, respectively. These games di¤er in two important
ways. First, they correspond to di¤erent ratios of the amount player B has to
sacri…ce in order to give player A a particular reward. According to the theory
discussed in the previous section, player B is only willing to give this reward if
his empathy parameter i exceeds threshold . The latter di¤ers between the
M -game and the two I-games, see the …nal column in Table 2. Second, the M
and I games also di¤er in the amount player A forgoes by choosing to enter.
The remaining six games have been chosen using games I through III as
starting point. Game IV combines the payo¤s of the ’stay-out’ option of the
M -game with the sacri…ce-reward values of the two I-games. Games V and VI
do so the other way around. The …nal three games take the stay-out payo¤s
as in games I through III and combine these with equal sacri…ce-reward values
such that an
of one half results.
We used the …rst part to generate a ’track record’for each individual. Such a
track record consisted of a two-tuple (ei ; ri ), with ei ; ri 2 f0; 1; :::; 9g the number
of enter choices subject i made as player A and the number of ’right’choices s/he
made as player B, respectively. Note that this track record does not indicate to
which of the nine games the ei enter choices and the ri right choices belong. In
that sense the track record only provides aggregate (or general) information.
1 6 First they made 9 decisions as player A, after that they made 9 (conditional) decisions as
player B. The games were presented in the same order as in Table 2.
10
After all subjects completed the …rst part, the second part started and subjects …rst learned their roles. They could either be an owner ("person C") or an
employee ("group member"). Subjects kept the same role throughout the entire
second part. Roles were assigned as follows. In each session we ranked the 45
subjects on the basis of the number of ’right’ choices ri in their track record.
Subjects with rank 16 to 30 were assigned the role of owner. The remainder of
the subjects were assigned the role of employee. This procedure –unknown to
the subjects – secured that we had enough variation in empathy types among
employees.
Once roles were assigned, subjects played 15 periods. At the beginning of
each period, …rms (called "groups" in the experiment) were exogenously formed.
A …rm consisted of one owner and two employees. The 15 periods were divided
into three blocks of …ve. We used a stranger design, in which each subject met
each other subject exactly once. Subjects were thus never confronted with the
same …rm member again and they were explicitly informed about that.
In the …rst …ve periods, either the IL (IH ) game or the M game was played
in isolation (see Table 1 above). In each of these periods the owner …rst decided,
on the basis of the observed track records, which role to assign to each of her
two employees. The owner thus observed (e1 ; r1 ) and (e2 ; r2 ) of her two assigned
employees, and so did the two employees of each other.17 One of the employees
should be assigned the role of player A, the other one should be given the role of
player B. After employees were assigned their roles, they played the game that
applied (either I or M ), making decisions for the role assigned. This determined
their period payo¤s, as given in the respective extensive form games. The owner
received a period payo¤ equal those of player A in the I-game and equal to those
of player B in the M -game. To easily remind owners about this fact, we labelled
the I-game as game “Azul" and the M -game as game “Blanco" (and we printed
these games against the corresponding background color). The complete setup
of the game was common knowledge, except for the payo¤s of the owner, which
were known to players C only.18
In periods 6 to 10 the other game was played (cf. Table 1). Again, in
each period …rms were exogenously formed and the owner …rst decided on role
assignment. After that, the two employees made their respective choices and
period payo¤s were obtained. In the …nal …ve periods (11 to 15) the owner …rst
chose which game to play, either IL (IH ) or M . Once a game had been chosen,
the order of decisions was as before.
Except for the decisions made within their own …rm in a given period, subjects did not get information on how the other subjects behaved in part 2.
Although they may have recorded the decisions made by previous …rm members in earlier periods, this information is of limited value because they would
never meet with the same other subject again. The observable records they
obtained from part 1 were thus the main clue they could use to predict the
1 7 In the instructions owners were explicitly informed that their two employees observed the
track record of each other.
1 8 We did so to ensure that employees’ decisions (i.e. those of players A and B) were not
guided by empathic feelings towards player C.
11
behavior of other subjects within their …rm. The track record of owners was
never made public. Therefore, for one third of the subjects the individual track
record remained private information throughout the experiment.
Payo¤s were determined in the following way. From the …rst part, one game
was selected at random. For this particular game subjects were then randomly
coupled in pairs and were randomly assigned roles.19 The individual payo¤s
that resulted from these pairings gave the earnings for the …rst part. To this
amount we added the overall payo¤s from the second part. The conversion rate
was such that 500 points in the experiment corresponded with 1 euro in money.
Apart from that, subjects received a show up fee equal to 7 euros.
The experiment was computerized using the z-tree programming package.
Subjects started with written instructions (for the …rst part) which were also
read aloud by the experimenter. At the end of the …rst part subjects received
new instructions for the second part. Before the second part started, subjects
played one practice period. After …nishing the second part subjects …lled in
a short questionnaire. Having completed this, the experimental points earned
were exchanged for money and subjects were paid individually and discreetly.
4
Results
In this section we present the …ndings of our experiment. We …rst describe the
distribution of individual ‘track records’ as generated by the choices subjects
made in part one. In the second subsection we look at how players A and
B behave in the three entry-reward games (M , IH and IL ) at hand. This
subsection also tests our …rst two hypotheses that: (i) B’s are more likely to
reward A’s entry when they have a higher r-value in their own track record and
(ii) A’s in turn are more likely to enter the higher the r-value of player B with
whom they are coupled. The …nal subsection looks at the organizational design
choices made by players C. It is tested whether they assign the role of player B
to the employee with the higher r-value in his track record and whether they
are more likely to choose the M -game when the latter r-value is higher.
4.1
Individual track records
In part one each subject makes nine entry decisions as player A and nine (conditional) choices between Left and Right as player B. From these 18 choices,
an individual track record (ei ; ri ) results. Table 3 gives the overall distribution
observed for the 180 subjects in our experiment.
On average subjects choose to enter 3:76 times as player A and as player B
they choose Right on average 2:83 times. As the frequency distribution makes
clear though, there is quite some variation across individual subjects. Most observations are scattered around the diagonal (see the numbers printed in bold),
1 9 Because we had an odd number of subjects within each session (45), we actually assigned
22 subjects the A-role and 23 subjects the B-role. The decision of one random A-subject was
then used twice to determine the payo¤s of two di¤erent B-roles.
12
Table 3: Overall distribution of individual track records
number of e choices
# of r’s
0
1
2
3
4
5
6 7 8 9 Total
0
10 11
9
4
4
3
1
42
1
1
2
4
5
5
4
2
23
2
2
5
1
7
1
2
18
3
2
5
3
6
4
3
3 1
27
4
3
6 13
5
4
2
33
5
1
2
3
4
6
1
17
6
1
1
2
1 1
6
7
1
2
1
4
8
1
1
2
9
2
1
1
1 3
8
Total
11 16 27 26 37 35 11 3 7 7
180
Remark: Numbers on the diagonal where ei = ri appear in bold.
suggesting that the number of Enter and Right choices are correlated. Indeed,
for our full sample of 180 subjects the Spearman rank correlation between e and
r choices equals 0:48 and is highly signi…cant (p = 0:000).20 Moreover, many
entries in Table 3 are above the diagonal. This indicates that subjects typically
choose as player A Enter somewhat more often than they choose Right as player
B, which is corroborated by formal signrank tests.21
There are some minor di¤erences in the observed track records across sessions. Comparing the number of r-choices by means of a Kruskal-Wallis test,
we do not …nd a signi…cant di¤erence (at the 5% level) between the four sessions (p = 0:0723). For the number of enter choices there are some di¤erences
though (p = 0:0131). As it appears, subjects in session 3 have a lower e-value
than those in sessions 1 and 4.22 In the former the average equals 2:87, in the
latter two 4:13 and 3:96, respectively. But even in session 3 the average e-value
(2:87) exceeds the average r-value (2:44). The main observation of a substantial
correlation between e and r-choices with (slightly) higher e-choices thus applies
to all sessions.
As explained in the previous section, after part one we ordered the subjects
on the basis of their number of r-choices. The middle third (ranks 16 to 30)
were assigned the role of player C. In sessions 1 and 4 these were subjects with
ri = 2 to ri = 4, whereas in sessions 2 and 3 these where subjects with ri = 1
to ri = 3 and with ri = 1 to ri = 4; respectively. Recall that in every period
players C were assigned one employee from the low-r group and another one
from the high-r group.
2 0 This also holds when we compute correlations for each of the four sessions in isolation
and when we compute standard (instead of Spearman rank) correlations.
2 1 Only in session 3 we do not …nd a signi…cant di¤erence (at the 5%-level) in indivivual e
and r scores. But for the three other sessions we do, as well as overall.
2 2 This follows from performing both ranksum and Kolmogorov-Smirnov tests.
13
4.2
Table 4: Number
Outcome Periods 1-5
Out
62 (83%)
E&L
7 (9%)
E&R
6 (8%)
Total
75 (100%)
of outcomes by
Periods 6-10
73 (97%)
1 (1%)
1 (1%)
75 (100%)
period for Game IL
Periods 11-15 Total
70 (95%)
205 (92%)
2 (3%)
10 (4%)
2 (3%)
9 (4%)
74 (100%)
224 (100%)
Table 5: Number
Outcome Periods 1-5
Out
55 (73%)
E&L
11 (15%)
E&R
9 (12%)
Total
75 (100%)
of outcomes by
Periods 6-10
61 (81%)
5 (7%)
9 (12%)
75 (100%)
period for Game IH
Periods 11-15 Total
30 (75%)
146 (77%)
6 (15%)
22 (12%)
4 (10%)
22 (12%)
40 (100%)
190 (100%)
Employees’choices
We next look at the decisions players A and B make in part 2 of the experiment.
Recall that each of the three di¤erent games M , IH and IL has the same decision
structure as in Figure 2. Tables 4 through 6 provide an overview of the outcomes
observed.
In periods 1 to 10 the game was exogenously given whereas in the last …ve
periods it was endogenously chosen by player C. The predominant outcome in
game IL is that player A chooses Out. In the very few instances that A chooses
to enter, player B is about equally likely to choose either Left or Right. The
latter also applies for game IH , but there player A is somewhat more likely to
enter. Finally, in the motivation game it is (much) more likely that player A
enters than in the two inspection games. But there player B also appears more
likely to choose Left over Right.
The …rst hypothesis we want to test is whether a subject with a higher rvalue in his/her track record is more likely to choose Right if assigned the role of
player B. For the three games together we have overall 237 cases in which player
A chooses to enter and thus B gets to make a choice. We use these observations
to estimate a random e¤ects probit model of the probability that B chooses
Right. Table 7 reports the results.
The four speci…cations reported all include player B’s r-value from his track
Table 6: Number
Outcome Periods 1-5
Out
84 (56%)
E&L
47 (31%)
E&R
19 (13%)
Total
150 (100%)
of outcomes by period for Game M
Periods 6-10 Periods 11-15 Total
98 (65%)
130 (70%)
312 (64%)
40 (27%)
47 (25%)
134 (28%)
12 (8%)
9 (5%)
40 (8%)
150 (100%)
186 (100%)
486 (100%)
14
Table 7: Random e¤ects probit estimations of B choosing Right
(1)
(2)
(3)
(4)
ri -own
0.111**
0.186**
0.105**
0.180*
(0.050)
(0.095)
(0.049)
(0.095)
IH -game
0.056
0.189
-0.113
0.068
(0.486)
(0.515)
(0.625)
(0.687)
M -game
-0.689
-0.612
-0.635
-0.478
(0.444)
(0.458)
(0.558)
(0.582)
period
-0.062**
-0.069**
-0.150*
-0.166*
(0.028)
(0.029)
(0.082)
(0.086)
ei -own
0.032
0.027
(0.093)
(0.092)
ei -other
-0.069
-0.088
(0.070)
(0.072)
ri -other
0.147*
0.152*
(0.089)
(0.091)
second
0.712
1.007
(1.407)
(1.457)
endo
1.179
1.487
(1.165)
(1.207)
IH second
0.408
0.171
(1.452)
(1.505)
M second
-0.233
-0.471
(1.397)
(1.442)
IH endo
-0.207
-0.410
(1.073)
(1.115)
M endo
-0.279
-0.538
(0.925)
(0.959)
constant
-0.177
-0.803
0.021
-0.575
(0.483)
(0.718)
(0.584)
(0.856)
Log L
N
rho
LR-chi2
-126.324
237
0.452***
24.607***
-124.673
237
0.490***
27.909***
-125.123
237
0.433***
27.008***
-123.324
237
0.472***
30.606***
Remark: Standard errors in parentheses.
= = indicates signi…cance
at the 1=5=10% level. Rho gives the proportion of overall variance
contributed by the panel-level component; its signi…cance is based on a
likelihood ratio test that rho=0. LR-chi2 reports the test statistic from
testing that all coe¢ cients (except the constant) are zero.
15
record. Apart from that, in the …rst column two 0=1-dummies for respectively
the IH -game and the M -game are incorporated (so IL serves as baseline), together with a time trend ‘period’. The second column adds the remaining variables of B’s and A’s track record. These two speci…cations do not distinguish
between whether the game played is the …rst or the second exogenous game
played in a row, or whether it is endogenously chosen by player C. Speci…cations (3) and (4) add additional zero-one dummies to identify potential order
and treatment e¤ects in this regard.23
In line with our …rst hypothesis, player B’s own r-value is signi…cant in all
speci…cations. Also the r-value of player A appears (marginally) signi…cant.
The latter suggests that player B’s reaction to A’s entry choice partially depends on how A would have behaved were he in B’s position. The e-values in
the track records of A and B do not have a signi…cant impact. Although Tables
4 through 6 above suggest that B is less likely to choose Right in the motivation game, the probit estimates for the M -game dummy are insigni…cant (but
do have the expected negative sign). The time trend (period) is signi…cantly
negative, indicating that the propensity to choose Right decreases over time.
We summarize our main …nding from Table 7 in Result 1.
Result 1. B’s with a higher r-value in their track record are more likely to
choose Right in all three games.
We next turn to A’s choice between Enter and Out. This choice is observed for
every interaction in part 2, so we have 900 observations in this case. Table 8
gives the random e¤ects probit estimates of the probability to enter.
In all speci…cations, the r-value in B’s track record (…rst row) is a highly
signi…cant determinant of A’s decision to enter. The higher rB is, the larger
A’s entry propensity. Entry is also signi…cant more likely in the IH and the
M -game as compared to the IL -game. Apart from that, also player A’s own
r-value increases his probability of choosing enter. The e-values in A’s and
B’s track record do not play a signi…cant role though. In particular, player A’s
own e-value does not provide much information about his probability of entry.24
Overall, these results are in line with our second hypothesis.
Result 2. The higher rB is, the more likely it is that player A enters.
Results 1 and 2 are consistent. The higher rB is, the more likely it is that
player B chooses Right. This in turn makes it more attractive for A to enter,
in line with what we observe. Player B’s track record thus contains valuable
2 3 The dummy variable ‘second’ equals one if period is in between 6 and 10, and zero otherwise. Variable ‘endo’ equals one i¤ period exceeds 10. The four interaction terms multiply
the game dummies with second and endo, respectively.
2 4 The insigni…cance of e -own in speci…cations (2) and (4) of Table 8 is only partly due to the
i
fairly substantial correlation between ei -own and ri -own (potentially leading to problems of
multi-collinearity). If we leave ri -own out of speci…cation (2), we obtain a coe¢ cient estimate
of 0:056 for ei -own with a p-value of 0:084. If we leave ri -own out of speci…cation (4), the
coe¢ cient for ei -own remains insigni…cant (p = 0:460).
16
Table 8: Random e¤ects
(1)
ri -other
0.105***
(0.022)
IH -game
0.702***
(0.197)
M -game
1.384***
(0.170)
period
-0.055***
(0.012)
ri -own
ei -own
ei -other
second
endo
IH second
M second
IH endo
M endo
constant
-1.559***
(0.199)
Log L
N
rho
LR-chi2
-451.243
900
0.187***
120.354***
probit estimations of A choosing Enter
(2)
(3)
(4)
0.176***
0.103***
0.158***
(0.033)
(0.023)
(0.044)
0.754***
0.548*
0.669*
(0.198)
(0.281)
(0.367)
1.364***
1.206***
0.680**
(0.172)
(0.249)
(0.316)
-0.056***
-0.134***
-0.139***
(0.012)
(0.036)
(0.051)
0.106***
0.094*
(0.037)
(0.052)
0.012
-0.009
(0.035)
(0.050)
-0.004
0.021
(0.031)
(0.043)
-0.169
0.110
(0.437)
(0.589)
0.679
0.991
(0.487)
(0.655)
0.353
0.205
(0.488)
(0.636)
0.517
0.238
(0.454)
(0.602)
0.491
-0.004
(0.449)
(0.578)
0.210
-0.273
(0.371)
(0.493)
-2.111***
-1.165***
-2.293***
(0.286)
(0.256)
(0.460)
-445.721
900
0.158***
131.399***
-445.301
900
0.193***
132.238***
-214.979
900
0.202***
50.572***
Remark: Standard errors in parentheses.
= = indicates signi…cance at
the 1=5=10% level. Rho gives the proportion of overall variance contributed
by the panel-level component; its signi…cance is based on a likelihood ratio
test that rho=0. LR-chi2 reports the test statistic from testing that all
coe¢ cients (except the constant) are zero.
17
Table 9: Assignment of roles in
rA > r B
rA < rB
eA > eB 72 (32%) 11 (5%)
eA = eB 9 (4%)
16 (7%)
eA < eB 8 (4%)
108 (48%)
Total
89 (40%) 135 (60%)
IL -game
Total
83 (37%)
25 (11%)
116 (52%)
224 (100%)
Table 10: Assignment of roles in
rA > r B
rA < rB
eA > eB 52 (27%) 6 (3%)
eA = eB 6 (3%)
10 (5%)
eA < eB 13 (6%)
103 (54%)
Total
71 (37%) 119 (63%)
IH -game
Total
58 (31%)
16 (8%)
116 (61%)
190 (100%)
information about how he is going to behave, which is actually used by A to
guide her entry decision. In the next subsection we investigate whether this
mechanism is recognized as such by player C when deciding on organizational
design.
Another important observation that follows from Tables 7 and 8 is that,
given the employees’ r-values, their e-values do not provide useful additional
information about their likely future behavior. Put di¤erently, the informational
value of the track records lies in the r-values; the e-values are informative only
to the extent that they are correlated with the r-values.
4.3
Allocation of roles and organizational design
In part two player C makes two types of choices. First, in every period she has
to decide on role assignment. Observing the track records (ei ; ri ) and (ej ; rj )
of her two employees, player C decides who gets the role of player A and who
becomes player B. Moreover, in periods 11 to 15 player C also chooses, before
role assignment, the game that is going to be played.
Tables 9 through 11 provide an overview of the assignment decisions in the
three di¤erent games. These tables reveal whether the employee who obtained
role B is the one with the higher r-value in his track record (rA < rB ) or whether
this is the other way around (rA > rB ), and similarly so for the e-values. Owing
to our role assignment procedure based on the ranking of ri -values, the two
employees within a group never had the same r-value. This does not apply
to the ei -values though, explaining the additional row where eA = eB in these
tables.
From the observed assignment patterns it immediately follows that our third
hypothesis is rejected; players C do not always assign the role of player B to
the employee with the higher r-value in his track record. In the two inspection
games it does hold that C is more likely to assign role B to the employee with the
18
Table 11: Assignment of roles in
rA > r B
rA < rB
eA > eB 281 (58%) 32 (7%)
eA = eB 37 (8%)
15 (3%)
eA < eB 11 (2%)
110 (23%)
Total
329 (68%) 157 (32%)
M -game
Total
313 (64%)
52 (11%)
121 (25%)
486 (100%)
higher r-value. In the motivation game this is actually the other way around;
there player C is more likely to assign role A to the higher r-value within his
work force.25 Given the fairly high correlation between subjects’r and e-values,
the order of e-values among the two employees typically corresponds with the
order of r-values. Assignment on the basis of relative e-values thus often (but
not always) coincides with allocation on the basis of relative r-values.
Comparing assignment patterns of r-values across games, we …nd that these
do not di¤er between the two inspection games. But they are signi…cantly
di¤erent for the motivation game; the average value of rB (per individual player
C) is lower in the motivation game as compared to the average value of rB (per
individual player C) in the inspection games.26 As a result of this, managers
in the M -game (players B) have on average an ri -value that does not di¤er
signi…cantly from managers in the I-game (players A).27 This …nding contrasts
with the theoretical prediction that managers in the M -game would be more
empathic than managers in the I-game.
Result 3. Players C quite often do not assign the role of player B to the
employee with the higher r-value. In the inspection games C is more likely to
assign role B to the high-r employee, in the motivation game she is more likely
to give role B to the low-r employee.
Result 3 is opposite to what we expected. Especially in the M -game it is
important for player C to stimulate entry (because the payo¤ after Out is low),
and assigning the high-r type to role B appears an e¤ective instrument to do so
(cf. Results 1 and 2). Yet the majority of player C’s does not do this.
A potential explanation why C’s in game M tend to assign role B to the low-r
employee is that they naively assume that A’s and B’s decisions are a¤ected only
2 5 To account for the multiple assignment decisions per player C, we formally test this as
follows. For each individual player C we compute the average values of rB and rA for each
game separately. We then compare these individual means r B and r A by means of signrank
tests. For the two I-games r B is signi…cantly larger than r A (p = 0:0472 in IL and p = 0:0045
in IH ), for the M -game r B is signi…cantly lower than rA (p = 0:0000).
2 6 These conclusions are based on comparing the means r
B of individual player C’s across
games. For IL versus M and IH versus M signrank tests (for matched pairs) yield p-values
of 0:0020 and 0:0015, respectively. For IL versus IH a ranksum test (unmatched data) gives
p = 0:9528.
2 7 This follows from comparing (per individual player C) r
B under game M with r A under
games IL and IH , respectively. Using signrank tests we obtain a p-value of 0:6288 for M
versus IL and of 0:9508 for M versus IH .
19
by their own eA and rB values, respectively. In particular, C’s may overlook
that A’s entry decision is mainly guided by the value of rB (cf. Result 2). If
C’s indeed have such naive expectations, they would prefer to assign the high-e
employee to role A and the low-r employee to role B. This follows because C gets
the same as player B in game M and thus is necessarily better o¤ when A enters
instead of staying out (and she is best o¤ when B chooses Left in reaction). This
could explain the direction in the assignment patterns we observe; high (ei ; ri )
types typically get role A whereas low (ei ; ri ) types usually get role B. Moreover,
one would expect naive C’s to focus predominantly on relative e-values in order
to stimulate entry; stimulating Left in reaction to entry is useful only when
entry can be induced.
For the two inspection games matters are less clear under naive expectations.
Surely, C then prefers to give role B to the high-r employee. This maximizes the
probability that B chooses Right after A enters. But given that C gets the same
as A in these games, entry is now risky for player C; she may end up with the
lowest payo¤ when B chooses Left upon A’s entry. It is therefore a priori unclear
whether a naive C wants to increase the probability of entry by assigning the
high-e employee to role A, or whether she prefers to avoid the worst outcome
(Enter, Left) in this game by giving the low-e employee role A. Under naive
expectations one therefore expects that allocation decisions may be driven by
both the relative r-values and the relative e-values of the two employees.
Given the substantial correlation between ei and ri , it is di¢ cult to identify
precisely the separate e¤ects of ei and ri on the probability of obtaining role B.
But the random e¤ects probit estimates in Table 12 provide suggestive evidence.
These estimates are calculated as follows. For each allocation decision of player
C we focus on the employee with the lower subject id in player C’s current
group of (two) employees. Because subject id’s are allocated at random, this
corresponds to a random selection of one of the two employees player C has in
a particular period. For these employees we estimate the probability that they
are assigned role B, where the random e¤ects procedure takes account of the
fact that we have multiple allocation decisions per player C within our sample.
We pool the data from the two inspection games, because as already discussed
above allocation patterns were the same in IL and IH . However, we report
separate estimates for the cases where the game at hand is exogenously …xed
and where it is endogenously chosen. We do so because allocation patterns may
be a¤ected by the determinants that drive the choice of organizational design.
Two explanatory variables are included: (i) the di¤erence in r-values between
the (lower id) employee and the other employee and (ii) the di¤erence in e-values
between them. For the two inspection games …ndings depend on whether the
game is exogenously given or whether it is endogenously chosen. In the former
case only the di¤erence in r-values appears signi…cant; this suggests that player
C’s particularly focus on assigning the higher r-value to role B. However, when
the inspection game is endogenously chosen only the di¤erence in e-values is
signi…cant. Then players C prefer to assign the higher e-value to role B and
thus the lower e-value to role A. A plausible explanation is that by doing so,
players C hope to reduce entry. The results for the motivation game are largely
20
Table 12: RE probit estimations of (lower id) employee getting role B
I-games
I-games
M -game
M -game
exo given endo chosen exo given endo chosen
ri -own ri -other 0.049**
0.032
-0.039*
-0.006
(0.019)
(0.034)
(0.021)
(0.027)
ei -own ei -other 0.011
0.119**
-0.086*** -0.190***
(0.029)
(0.054)
(0.032)
(0.043)
constant
0.001
-0.144
-0.000
0.119
(0.078)
(0.137)
(0.098)
(0.108)
Log L
N
rho
LR-chi2
-200.506
300
0.022
15.009***
-70.560
114
0.039
16.831***
-189.280
300
0.153**
34.062***
-103.603
186
0.029
50.268***
Remark: Standard errors in parentheses.
= = indicates signi…cance at the
1=5=10% level. ‘exo given’ refers to periods 1 to 10 where the game at hand is
exogenously given, ‘endo chosen’ refers to periods 11 to 15 where the game is
chosen by player C as well.
independent of whether this game is given or endogenously chosen. In this game
C’s are mostly concerned with getting the low-e value in role B, that is, assigning
the high-e value to role A.
These patterns makes sense when C’s have naive expectations as described
earlier. In the two inspection games C’s will be mostly concerned with avoiding
the very unattractive outcome (Enter,Left), and they think they can do so
either by having the higher r-value in role B and/or the lower e-value in role A.
The shift in focus towards the latter when the I-game is endogenously chosen
is in line with the relative e-values being decisive in actual game choice (see
below).28 In the motivation game players C particularly would like to stimulate
entry. Naive C’s think that this is best accomplished by assigning the higher
e-value to role A.
The above discussion suggests that, due to their naive expectations, players C make suboptimal allocation decisions, especially in the motivation game.
Result 2 namely indicates that entry is best stimulated by assigning the higher
r-value to role B.29 On the other hand, such an allocation also stimulates B to
2 8 Note, however, that owing to the substantial correlation between e and r , assigning
i
i
the higher r-value to role B typically coincides with assigning the lower e-value to role A.
Therefore, for many allocation decisions this shift in focus does not have an impact.
2 9 Result 2 is based on the probit estimates of A choosing Enter in Table 8. In this table r i
other (ei -other) corresponds with rB (eB ) and ri -own (ei -own) with rA (eA ). These estimates
take players A as the unit of analysis, which is appropriate given the focus there on explaning
A’s entry decision. If we focus on assignment decisions of players C, however, the cross
sectional units should correspond to di¤erent players C. We therefore re-estimated the probits
using players C as the clustering variable and we basically get the same results as in Table 8;
the estimated coe¢ cient of rB is signi…cantly larger than the one of rA . So entry is indeed
21
Table 13: Random e¤ects regressions of player C’s pro…t
IL -game
IH -game
M -game
rB
-4.553
-6.728
22.406***
(3.348)
(5.769)
(5.683)
rA
-8.558**
-0.492
13.064**
(3.838)
(6.682)
(5.388)
eA
0.217
-2.464
5.044
(2.896)
(6.393)
(4.824)
eB
-3.808
8.310
-0.113
(2.614)
(6.361)
(5.212)
second
37.452
27.195
93.278**
(25.307)
(39.214)
(40.983)
endo
60.620
5.393
183.310***
(41.581)
(62.115)
(64.670)
period
-4.429
-1.211
-24.350***
(3.862)
(5.789)
(6.128)
constant
479.387*** 336.771*** 408.180***
(30.484)
(39.035)
(43.499)
Overall R2
N
rho
Wald-chi2
0.056
224
0.085
16.521**
0.031
190
0.145
9.026
0.063
486
0.150
35.841***
Remark: Standard errors in parentheses.
= = indicates signi…cance at the 1=5=10% level. Rho gives the proportion of
overall variance contributed by the panel-level component. In
all three speci…cations a Lagrange multiplier test for random effects is insigni…cant. Wald-chi2 reports the test statistic from
testing that all coe¢ cients (except the constant) are zero.
choose Right after Enter, and (conditional on entry) C would be better o¤ if B
chooses Left instead. To assess the overall e¤ect we therefore investigate how
player C’s pro…ts vary with her allocation decision. Table 13 presents random
e¤ects regression estimates of player C’s pro…t for each of the three games separately. As explanatory variables we include the track record characteristics of
C’s two employees, together with two treatment dummies and a time trend.
First focussing on the motivation game, we observe that the e-values of
player A and B do not signi…cantly a¤ect player C’s pro…ts. The r-values, on
the other hand, signi…cantly increase pro…ts. It also holds that the coe¢ cient of
rB is signi…cantly larger than the one of rA (p = 0:0537). Player C would thus
make more pro…t if she would give the employee with the higher r-value role B.
Because she typically does not do so (cf. Result 3), allocation decisions in the
best stimulated by assigning role B to the high r-value in the work force.
22
Table 14: Choices between games by session
Session Game IL Game IH Game M
Total
1
39 (52%) 36 (48%)
75 (100%)
2
35 (47%) 40 (53%)
75 (100%)
3
24 (32%) 51 (63%)
75 (100%)
4
16 (21%) 59 (79%)
75 (100%)
Total
74 (25%) 40 (13%) 186 (62%) 300 (100%)
motivation game are indeed suboptimal.
In regard to the two inspection games no signi…cant coe¢ cients are found for
the employees’track record variables. An explanation for this is that a higher
rB -value not only makes outcome (Enter, Right) more likely, but also the worst
possible outcome (Enter, Left).30 The regression results suggest that the payo¤
consequences of these two opposing e¤ects cancel out. Pro…t levels are thus
largely insensitive to assignment and therefore the allocation decisions cannot
be labelled suboptimal for the inspection games.
Result 4. In the motivation game Players C make more pro…t when they assign
the employee with the higher r-value to role B. In the inspection games pro…t
levels are largely insensitive to the assignment of the employees.
We …nally look at the choice C makes between the I-game and the I-game in
the …nal …ve periods of part 2. From Table 14 it can be observed that when
player C chooses between IL and M , she is about equally likely to choose either
game. But when the choice is between IH and M , she chooses the motivation
game in overall almost 75% of the cases. In line with theoretical predictions,
therefore, C is more likely to choose M over IH than M over IL .
To explore game choice in more detail, Table 15 reports random e¤ects probit
estimates of the probability that game M is chosen. The …rst two columns pool
the choices made between IL and M and between IH and M , respectively. Here
rHigh (rLow ) refers to the r-value of the employee with the higher (lower) r in
his track record. Variables eHigh and eLow are de…ned similarly. Apart from the
track record characteristics of player C’s two employees, the …rst speci…cation
also includes a dummy equal to one i¤ IH is the alternative, a time trend, and
a variable M
I measuring the di¤erence in average realized pro…ts player C
obtained from the two games in the …rst ten periods (where these games were
exogenously given). Intuitively, one would expect that C is more likely to choose
game M when this game yielded her higher pro…ts than game I did in the past.
In the second speci…cation we include player C’s own track record as well. The
…nal two columns consider the two subsamples M versus IL and M versus IH
in isolation.
3 0 This follows from running separate RE probit estimates of the probability of outcome
(Enter,Left) and (Enter,Right), respectively.
23
Table 15: Random e¤ects probit estimations of C choosing game M
(1)
(2)
(3)
(4)
M vs. I
M vs. I
M vs. IL M vs. IH
rHigh
0.048
0.053
0.086
-0.002
(0.054)
(0.056)
(0.073)
(0.111)
rLow
0.180
0.216
0.350*
-0.089
(0.144)
(0.153)
(0.182)
(0.293)
eLow
-0.013
-0.011
-0.104
0.174
(0.063)
(0.062)
(0.080)
(0.136)
eHigh
0.166***
0.165***
0.201***
0.093
(0.055)
(0.055)
(0.061)
(0.129)
order
0.100
0.046
0.211
0.285
(0.216)
(0.226)
(0.380)
(0.518)
M versus IH 0.719***
0.727***
(0.242)
(0.242)
0.003***
0.003***
0.004***
0.004
M
I
(0.001)
(0.001)
(0.001)
(0.003)
period
-0.038
-0.039
-0.112
0.057
(0.057)
(0.057)
(0.078)
(0.092)
ri -own
-0.075
0.044
-0.265
(0.118)
(0.149)
(0.303)
ei -own
0.023
0.035
0.028
(0.060)
(0.067)
(0.132)
constant
-0.875
-0.781
-0.523
-0.432
(0.850)
(0.930)
(1.300)
(1.710)
Log L
N
rho
LR-chi2
-174.725
300
0.175**
30.153***
-174.420
300
0.172**
30.723***
-90.189
150
0.009
23.389***
-75.674
150
0.439***
7.148
Remark: Standard errors in parentheses.
= = indicates signi…cance
at the 1=5=10% level. Rho gives the proportion of overall variance
contributed by the panel-level component; its signi…cance is based on a
likelihood ratio test that rho=0. LR-chi2 reports the test statistic from
testing that all coe¢ cients (except the constant) are zero.
24
In all four speci…cations variable rHigh is far from signi…cant. The estimates
thus do not lend support for the hypothesis that the higher the value of rHigh
within player C’s work force is, the more likely it is that the M -game is chosen.
The single variable from the employees’track records that appears a signi…cant
determinant (in 3 out of 4 speci…cations) is eHigh . The higher eHigh is, the more
likely it becomes that player C chooses M: The only two other variables that
attain signi…cance are the pro…t di¤erence M
I and the M versus IH dummy.
Not surprisingly, the better (relative) experience player C has with game M in
the past, the more likely she is to choose this game over the inspection game.
And similar so the worse the alternative is, i.e. IH instead of IL .
Result 5. (i) Players C’s choice between the motivation game and the inspection game is not guided by the ri -values in her employees’track records. From
these records only the highest e-value eHigh matters; the higher eHigh , the more
likely it is that C chooses game M . (ii) player C is more likely to choose M over
IH than over IL :
Overall the following general picture emerges. For players C the motivation
game is attractive only if entry can be induced in this game. They naively think
that this can be best accomplished by allocating role A to the employee with
the highest ei -value in the work force. Player Cs then choose game M when
this eHigh -value is relatively high. In case eHigh is low Cs are more likely to
choose the inspection game and will assign this player the role of B. A rationale
for the latter assignment is that player C hopes to avoid the very bad outcome
(Enter,Left) in this way. Allocation decisions are naive in the sense that Cs
seem to overlook the signi…cant impact of B’s track record (especially rB ) on
A’s entry choices, especially in game M where this relationship is of vital importance. They therefore mainly look at an employee’s own track record to form
expectations about how he would behave in role A. This leads to suboptimal
allocation choices in game M . In turn, the choice between the two games is
distorted as well.
5
Conclusion
Organizations di¤er widely in the practices they use to motivate their employees. Some organizations heavily rely on formal contracts with explicit incentives
and active monitoring. One important task of managers in such organizations
is to supervise and inspect workers in order to detect potential shirking. Other
organizational modes are predominantly based on implicit informal agreements
that hard work will be rewarded. In this case the main task of managers is to
inspire and to motivate the work force. In these organizations employment contracts are largely incomplete and a substitute mechanism is needed to convince
workers that the organization is indeed committed to reward high e¤ort. Apart
from repeated interaction and reputation (cf. Kreps (1990)), appointing managers that empathize with their employees may provide such a commitment (cf.
25
Rotemberg and Saloner (1993)). This motivational mode where managers with
social preferences are hired saves on the costs of using a formal performance
measurement system.
In this paper we test several predictions concerning organizational design, by
means of a laboratory experiment. Theory predicts that the motivation-mode
is viable only if managers are su¢ ciently empathic whereas for the inspectionmode this is not the case. The more empathic employees within the work force
should therefore be given the managerial positions in the M structure (but not
in the I-mode). And the more empathic these managers are, the more attractive
organizational mode M becomes relative to the inspection mode.
Our main …ndings are that owners by and large overlook the signi…cant impact of a manager’s preference type on worker behavior in the M -mode. They
naively assume that the worker’s e¤ort decision is mainly guided by the preference type of the worker himself. They therefore allocate roles suboptimally
in the M -mode, with workers rather than managers being the more empathic
types. As a result of this, choices between organizational modes di¤er from theoretical predictions as well. Overall we conclude that owners in our experiment
do not make e¤ective use of the available preference types within their work
force when drafting their organizational design.
Appendix: basic model of endogenous organizational design
In the experiment subjects are confronted with the simple games depicted in
Figure 1. To motivate the particular parameter values we have chosen, we
consider in this Appendix a bare bone (reduced form) model of endogenous
organizational design. Because the main purpose here is to justify our parameter
choices, in this model we abstract away from the owner’s assignment decision.
A …rm consists of three agents: the owner who owns the …rm, a manager
hired to run the …rm on her behalf and a worker doing the productive work.
The worker can either put in low e¤ort (‘shirk’) or high e¤ort (‘work’). In the
former case the value of his productivity equals v0 whereas in the latter case it
is v1 .31 The worker’s disutility of putting in high (instead of low) e¤ort equals
g. Therefore, a sel…sh worker will shirk if no additional measures are taken.
One way to motivate the worker to put in high e¤ort is to set up a performance monitoring system. We assume that such a system, when fully implemented, always induces the worker to work. It brings about three types of costs
though. First, there are the costs k of setting up and installing the monitoring
technology. Investments in technological equipment and organizational procedures are needed to allow accurate measurement of the worker’s productivity.32
Second, h denotes the …rm’s inspection costs. Even with the monitoring technology in place, scarce resources like the manager’s time need to be devoted to
3 1 With
respect to all parameters that will be introduced we assume that they are positive.
costs are equivalent to the investments in veri…cation technology required under
explicit contracts in the experiments of Fehr and Schmidt (2000) and Fehr et al. (2007).
3 2 These
26
monitor the worker. Third, the worker dislikes being monitored because it gives
him the feeling of being controlled, leading to a disutility of d. We assume that
the overall costs of the formal monitoring system fall short of the net bene…ts
of getting the worker to work:
k + h + d < v1
v0
g
(A1)
Therefore, in the absence of alternative incentive instruments, the …rm would
bene…t from using a formal monitoring system. As in in the main text we
will refer to this as the ‘inspection mode’ of organizational design, or I-mode
in short. The main task of the manager in this mode is to check whether the
worker does not deliver substandard work. In regard to compensation we assume
that the worker receives a …xed wage wI . The manager is paid on the basis of
performance pay, getting a share fI 2 (0; 1) of the …rm’s net pro…ts (while the
owner gets the remainder).
An alternative way to motivate the worker is to promise him a bonus whenever he puts in high e¤ort. Because e¤ort itself is non-contractable, this bonus
payment cannot be made part of a formal contract though. The incentive system thus relies on an implicit contract that the promise will be kept. This type
of organizational design is labelled as the motivation mode, or M -mode in short.
Here the main task of the manager is to inspire and to motivate, by developing
and maintaining a culture that high e¤ort is indeed rewarded by the …rm. In
the M -mode the worker receives a wage wM and is promised a bonus bM on top
of that if he exerts high e¤ort. The manager gets a fraction fM 2 (0; 1) of …rm
pro…ts. This performance pay gives a sel…sh manager an incentive to renege on
the promised bonus payment.
Overall the game model of Figure A1 results. First the owner chooses the
organizational mode. If the M -mode is chosen, the worker moves next by deciding whether to shirk or the work. Only if the worker works, the manager decides
whether to pay the promised bonus or not. (Here the implicit assumption is
that the manager never wants to reward shirking with a bonus.) In the I-mode
the manager moves before the worker does. The manager either commits to
monitor or not to do so. In the former case the worker is assumed to work,
because the disutility of working falls short of the costs of getting caught shirking. If the manager does not monitor, the worker chooses between shirking or
working. The players’payo¤s then follow from the assumptions made above.
< Insert Figure A1 about here >
If players are sel…sh, the predicted outcome is easily determined by backwards induction. A sel…sh manager will not pay the bonus in the M -mode
(given fM bM > 0). Anticipating this, a sel…sh worker will shirk under this
organizational design. In the inspection mode a sel…sh worker will shirk if not
monitored by the manager, therefore the manager will monitor him.33 The outcome is that the worker does work under this mode, at the expense of the overall
3 3 Note
that fI (v1
wI
k
h) > fI (v0
27
wI
k) follows from (A1).
costs of the monitoring technology (k + h + d). Given assumption (A1), under
sel…sh preferences the I-mode is more e¢ cient than the M -mode is. Hence if
the payo¤ parameters are such that the owner shares in these e¢ ciency gains,
she would choose the I-mode over the M -mode.34 It would be more e¢ cient,
however, if the worker could be motivated to work in the M -mode, as this would
save the overall costs of the monitoring technology.
Of course, in a fully ‡edged model the compensation parameters wM ; wI ; bM ;
fM and fI would be endogenous. Here we just make the following simplifying
assumptions. First, mainly for practical reasons we focus on the case in which
fM = fI = 12 .35 The share fraction is thus the same for the two organizational
modes, such that the owner does not simply prefer one mode over the other
because she can pay the manager less. Second, with respect to the wage and
bonus payments wM ; wI and bM we assume that:
wM =
fM v0
fI (v1 k h) + g + d
fM (v1 v0 ) + g
; wI =
; and bM =
1 + fM
1 + fI
1 + fM
The wage level wM (= v0 =3) ensures that all …rm members earn the same when
the worker chooses to shirk in the M -mode. Similarly so, wI is set such that all
…rm members get the same when the manager monitors in the I-mode. Finally,
bM makes that all members earn the same when the manager pays the bonus
in the M -mode after the worker decided to work. E¤ectively, payo¤ di¤erences
are minimized in the three most relevant outcomes and potential e¢ ciency gains
are shared equally.
Under these assumptions, the resulting payo¤s in the two modes follow directly from the (exogenous) production technology parameters appearing in inequality (A1). The payo¤s appearing in Figure 1 now result from making the
following choices:
v0 = 840; v1 = 1740; g = 90; k = 180; h = 260 [100] and d = 130 [50]
where for h and d the …rst value refers to IH and the second to IL .
References
Akerlof, G. and R. Kranton (2005). Indentity and the economics of organization. Journal of Economic Perspectives 19, 9–32.
Avenhaus, R., B. von Stengel, and S. Zamir (2002). Inspection games. In
R. Aumann and S. Hart (Eds.), Hanbook of Game Theory, Volume 3,
Chapter 51, pp. 1947–1987. North-Holland.
Calvo, G. and S. Wellisz (1978). Supervision, loss of control, and the optimum
size of the …rm. Journal of Political Economy 86, 943–952.
3 4 In
particular this requires (1 fM ) (v0 wM ) < (1 fI ) (v1 wI k h).
share fraction of one half is convenient in the experiment, because the owner’s (i.e.
player C’s) earnings then simply correspond to those of the second mover (player B) in the
M -mode and the …rst mover (player A) in the I-mode.
35 A
28
Charness, G. and M. Rabin (2002). Understanding social preferences with
simple tests. Quarterly Journal of Economics 117, 817–869.
Dickinson, D. and M.-C. Villeval (2004). Does monitoring decrease work effort? the complementarity between agency and crowding-out theories.
IZA Discussion Paper, No. 1222.
Dufwenberg, M. and G. Kirchsteiger (2004). A theory of sequential reciprocity. Games and Economic Behavior 47, 268–298.
Fehr, E., S. Gächter, and G. Kirchsteiger (1997). Reciprocity as a contract
enforcement device: experimental evidence. Econometrica 65, 833–860.
Fehr, E., A. Klein, and K. Schmidt (2007). Fairness and contract design.
Econometrica 75, 121–154.
Fehr, E. and J. List (2004). The hidden costs and returns of incentives trust and trustworthiness among ceos. Journal of the European Economic
Association 2, 743–771.
Fehr, E. and K. Schmidt (1999). A theory of fairness, competition, and cooperation. Quarterly Journal of Economics 114, 817–868.
Fehr, E. and K. Schmidt (2000). Fairness, incentives and contractual choices.
European Economic Review 44, 1057–1068.
Fehr, E. and K. M. Schmidt (2004). Fairness and incentives in a multi-task
principal agent model. Scandinavian Journal of Economics 106, 453–474.
Hermalin, B. (2001). Economics and corporate culture. In S. C. et al. (Ed.),
The International Handbook of Organizational Culture and Climate. John
Wiley and Sons Ltd.
Kreps, D. (1990). Corporate culture and economic theory. In J. Alt and
K. Shepsle (Eds.), Perspectives on Positive Political Economy. Cambridge
University Press.
Rabin, M. (1993). Incorporating fairness into game theory and economics.
American Economic Review 83, 1280–1302.
Rotemberg, J. (1994). Human relations in the workplace. Journal of Political
Economy 102, 684–717.
Rotemberg, J. and G. Saloner (1993). Leadership style and incentives. Management Science 39, 1299–1318.
Rotemberg, J. and G. Saloner (2000). Visionaries, managers and strategic
direction. Rand Journal of Economics 31, 693–716.
Schweitzer, M. and T. Ho (2005). Trust and verify: Monitoring in interdependent relationships. In J. Morgan (Ed.), Experimental and Behavioral
Economics, Volume 13.
29
Figure 1a. Motivation game M
Figure 1b: Inspection game IH [IL]
Worker
Manager
Shirk
Work
Monitor
Not Monitor
Manager
280
280
No Bonus
190
730
Bonus
550
550
Worker
360 [440]
360 [440]
Shirk
40
580
Work
490
490
Figure 2. General decision structure
(with d > c > b > a and d+a ≤ 2c)
Player A
Stay Out
Enter
Player B
b
b
Left
a
d
Right
c
c
Figure A1. The basic (reduced form) game
Owner
M-mode
I-mode
Worker
Shirk
Manager
Work
Monitor
Not Monitor
Manager
πO: (1−fM)⋅(v0−wM)
πW: wM
πMan: fM⋅(v0−wM)
No Bonus
πO: (1−fM)⋅(v1−wM)
πW: wM−g
πMan: fM⋅(v1−wM)
Legend:
πi: monetary payoffs for player i∈{O,Man,W};
v1 (v0) = value productivity if worker works (shirks);
g = worker’s cost of effort;
k = costs of setting up monitoring technology;
h = firm’s costs of inspection;
Bonus
Worker
πO: (1−fI)⋅(v1−wI−k−h)]
πMan: fI⋅(v1−wI−k−h)
πW: wI−g−d
πO: (1−fM)⋅(v1−wM−bM)
πW: wM+bM−g
πMan: fM⋅(v1−wM−bM)
Shirk
πO: (1−fI)⋅(v0−wI−k)
πMan: fI⋅(v0−wI−k)
πW: wI
d = worker’s disutility of being monitored;
wM (wI) = wage worker in M (I) mode;
bM = size of bonus payment in M-mode;
fM (fI) = profit sharing fraction manager in M (I) mode.
Work
πO: (1−fI)⋅(v1−wI−k)
πMan: fI⋅(v1−wI−k)
πW: wI−g